why flip twice?

[lxy5513]

Thanks for you great work!
read the code, I find some code which I can't to understand, as below,

why it need to flip append loc and append conf?

---

by the way, we compare loc and flip loc in

consistency_loc_loss_x = torch.mean(torch.pow(loc_sampled[:, 0] + loc_sampled_flip[:, 0], exponent=2))
...

but, how we guarantee each item between loc_sampled and loc_sampled_flip can match correctly?

Hope your kindly response, much thanks.

[soo89]

Thank you for your interest.

f^(p,r,c,d)(cls)(I) should match with f^(p,r,C-c+1,d)(cls)(I).
'L160 : append_conf = flip(append_conf,2)' is for matching above mentioned.

In our paper, we adopted not only classification loss but also regression loss.

Therefore, we flipped loc outputs.

f^(p,r,c,d)(loc)(I) should match with f^(p,r,C-c+1,d)(loc)(I).
'L159 : append_loc = flip(append_loc,2)' is for matching above mentioned.